Traditional control systems operate like a driver strictly committed to staying in the exact center of a lane. In the language of control theory, this is setpoint tracking: the controller minimizes a quadratic, least-squares cost function representing the system's deviation from a pre-calculated reference path.
However, in many real-world applications, the true goal is not to follow a arbitrary geometric path, but to maximize an abstract performance metric—such as financial profit, fuel efficiency, or energy output. Forcing a system to track a static reference is often an unnecessary constraint. In fact, as Gros demonstrates mathematically in An economic NMPC formulation for wind turbine control, traditional tracking objectives are merely second-order Taylor expansions of direct economic objectives. Standard industry tracking practices are, at best, approximations of the true goal.
Economic Model Predictive Control (EMPC) discards the surrogate tracking layer. Instead of minimizing deviation from a setpoint, it embeds the high-level economic objective directly into the predictive control loop over a receding time horizon. This shift allows controllers to proactively adjust behavior based on time-varying factors like fluctuating electricity prices, changing resource costs, or environmental disturbances.
The Stability Dilemma and the Turnpike Effect
If direct economic optimization is inherently more optimal, why has tracking dominated control theory for decades? The answer lies in stability guarantees.
When a controller minimizes the distance to a fixed coordinate, the system naturally settles into a steady state. But when a controller optimizes an arbitrary, indefinite economic metric, the most "efficient" behavior might involve continuous, unpredictable oscillation rather than settling down (Köhler et al.).
A classic manifestation of this instability is the turnpike effect, where the controller behaves greedily near the end of its look-ahead window. For example, an economic wind turbine controller trying to maximize power generation might aggressively slow down the rotor at the very end of its prediction horizon to extract a final burst of kinetic energy, leaving the system in an unviable state for the next time step. To eliminate this behavior, Gros introduced a terminal cost function that explicitly accounts for the rotor's remaining kinetic energy, forcing the controller to maintain long-term operational viability.
Tracking MPC: [Current State] ---> (Minimize Deviation) ---> [Fixed Setpoint]
Economic MPC: [Current State] ---> (Maximize Profit/ROI) ---> [Dynamic Trajectory]
|
+---> Danger: Greedy boundary behavior
(Mitigated by Terminal Costs/Storage Functions)
Proving that an EMPC system will remain stable under arbitrary economic objectives typically requires verifying a property called dissipativity and constructing a mathematical storage function. For complex, non-polynomial nonlinear systems, finding this storage function analytically is incredibly difficult. To bypass this bottleneck, Kordabad in Verification of Dissipativity and Evaluation of Storage Function in Economic Nonlinear MPC using Q-Learning framed the search as an undiscounted Q-learning problem, using reinforcement learning to automatically verify system stability and discover optimal control policies.
Furthermore, real-world systems must maintain these stability and feasibility guarantees even when subjected to random, non-periodic economic fluctuations or model inaccuracies. This has driven the development of stochastic frameworks for mode-switching operations (Sopasakis et al.) and adaptive learning mechanisms that update system parameters on the fly without violating performance bounds (Degner et al.).
Overcoming the Computational Bottleneck
The mathematical flexibility of EMPC comes at a steep computational cost. Quadratic tracking costs yield clean, convex optimization problems that can be solved almost instantly. General economic costs—such as maximizing the profit of a multi-fuel power plant (Kragelund et al.) or minimizing municipal water pumping costs (Wang et al.)—introduce complex, non-least-squares mathematical structures that traditional real-time solvers struggle to process quickly.
To make EMPC viable for fast, real-time industrial deployment, researchers focus on specialized optimization algorithms and structural decompositions:
- Algorithmic Acceleration: Hours et al. developed a parametric augmented Lagrangian algorithm designed specifically for non-least-squares economic costs. Their approach bounds the controller's sub-optimality over time while achieving microsecond-level solve times.
- Structural Decomposition: For resource-constrained hardware, such as Autonomous Underwater Vehicles (AUVs), Yang et al. introduced a two-stage motion decomposition that approximates the long-term "energy-to-go." This approach matched the efficiency of heavy global optimization while slashing computation time by 98%.
- System Linearization: Other approaches focus on simplifying the underlying system dynamics. For instance, Li et al. proposed combining Koopman operator theory with sequence-modeling architectures to linearize complex, time-varying systems, paving the way for more efficient predictive control under non-quadratic costs.
Emergent Behavior and the Safety Trade-off
When a controller is freed from the constraints of a fixed reference path, it can discover novel, highly efficient operational strategies that human engineers might not anticipate.
This phenomenon is clearly visible in robotic locomotion. In Economic model predictive control for snake robot locomotion, Nonhoff et al. used an EMPC framework to optimize for velocity and energy efficiency directly, rather than forcing a snake robot to track a pre-defined biological gait. The robot independently "discovered" a non-sinusoidal movement pattern that was faster and more energy-efficient than standard biological surrogates.
Traditional Tracking: Robot ---> [Forces adherence to standard sinusoidal gait]
Economic Discovery: Robot ---> [Optimizes velocity/energy directly] ---> Emergent non-sinusoidal gait
In transit and navigation, this flexibility manifests as an intelligent trade-off between path accuracy and resource consumption. By treating navigation errors as economic penalties rather than hard constraints, autonomous vehicles can choose to deviate slightly from a straight line to save massive amounts of fuel or battery life when fighting external disturbances. Yang et al. demonstrated that prioritizing cost minimization over rigid path-following can yield energy savings exceeding 50% in autonomous transits. Similarly, Deng et al. showed that treating tracking errors as energy penalties allows Autonomous Surface Vehicles (ASVs) to achieve significantly better tracking accuracy under ocean disturbances with negligible increases in power consumption.
However, direct economic optimization cannot happen in a vacuum; it must be balanced against physical safety and equipment longevity. Purely maximizing power capture in wind or marine current turbines can cause severe structural fatigue and actuator wear due to aggressive control actions during rapid transients. Modern EMPC frameworks resolve this by explicitly balancing immediate economic returns with mechanical degradation costs (Vukov et al.; Teng et al.). By integrating Lyapunov stability theory and set invariance directly into the distributed control loop, these systems guarantee that the pursuit of economic efficiency will never push critical infrastructure past safe physical boundaries (Cortes et al.).
Key Foundational Work
Predictive control methods to improve energy efficiency and reduce demand in buildings (2013) — contributed 1 concept elements to this direction.
Nonconvex model predictive control for commercial refrigeration (2013) — contributed 1 concept elements to this direction.
A Lyapunov function for periodic economic optimizing Model Predictive Control (2013) — contributed 1 concept elements to this direction.